arXiv subjects

Bowen Shi

Publications and source records attributed to Bowen Shi.

98 records · Page 6Linked to original sources

arXiv2019

Compression of Acoustic Event Detection Models with Low-rank Matrix Factorization and Quantization Training

In this paper, we present a compression approach based on the combination of low-rank matrix factorization and quantization training, to reduce complexity for neural network based acoustic event detection (AED) models. Our experimental results show this combined compression approach is very effective. For a three-layer long short-term memory (LSTM) based AED model, the original model size can be reduced to 1% with negligible loss of accuracy. Our approach enables the feasibility of deploying AED for resource-constraint applications.

eess.AS↗

arXiv2019

Semi-supervised Acoustic Event Detection based on tri-training

This paper presents our work of training acoustic event detection (AED) models using unlabeled dataset. Recent acoustic event detectors are based on large-scale neural networks, which are typically trained with huge amounts of labeled data. Labels for acoustic events are expensive to obtain, and relevant acoustic event audios can be limited, especially for rare events. In this paper we leverage an Internet-scale unlabeled dataset with potential domain shift to improve the detection of acoustic events. Based on the classic tri-training approach, our proposed method shows accuracy improvement over both the supervised training baseline, and semisupervised self-training set-up, in all pre-defined acoustic event detection tasks. As our approach relies on ensemble models, we further show the improvements can be distilled to a single model via knowledge distillation, with the resulting single student model maintaining high accuracy of teacher ensemble models.

eess.AS↗

arXiv2019

Multitask training with unlabeled data for end-to-end sign language fingerspelling recognition

We address the problem of automatic American Sign Language fingerspelling recognition from video. Prior work has largely relied on frame-level labels, hand-crafted features, or other constraints, and has been hampered by the scarcity of data for this task. We introduce a model for fingerspelling recognition that addresses these issues. The model consists of an auto-encoder-based feature extractor and an attention-based neural encoder-decoder, which are trained jointly. The model receives a sequence of image frames and outputs the fingerspelled word, without relying on any frame-level training labels or hand-crafted features. In addition, the auto-encoder subcomponent makes it possible to leverage unlabeled data to improve the feature learning. The model achieves 11.6% and 4.4% absolute letter accuracy improvement respectively in signer-independent and signer-adapted fingerspelling recognition over previous approaches that required frame-level training labels.

cs.CL↗

arXiv2019

American Sign Language fingerspelling recognition in the wild

We address the problem of American Sign Language fingerspelling recognition in the wild, using videos collected from websites. We introduce the largest data set available so far for the problem of fingerspelling recognition, and the first using naturally occurring video data. Using this data set, we present the first attempt to recognize fingerspelling sequences in this challenging setting. Unlike prior work, our video data is extremely challenging due to low frame rates and visual variability. To tackle the visual challenges, we train a special-purpose signing hand detector using a small subset of our data. Given the hand detector output, a sequence model decodes the hypothesized fingerspelled letter sequence. For the sequence model, we explore attention-based recurrent encoder-decoders and CTC-based approaches. As the first attempt at fingerspelling recognition in the wild, this work is intended to serve as a baseline for future work on sign language recognition in realistic conditions. We find that, as expected, letter error rates are much higher than in previous work on more controlled data, and we analyze the sources of error and effects of model variants.

cs.CV↗

arXiv2019

Characterizing topological order by the information convex

Motivated by previous efforts in detecting topological orders from the ground state(s) wave function, we introduce a new quantum information tool, coined the information convex, to capture the bulk and boundary topological excitations of a 2D topological order. Defined as a set of reduced density matrices that minimizes the energy in a subsystem, the information convex encodes not only the bulk anyons but also the gapped boundaries of 2D topological orders. Using untwisted gapped boundaries of non-Abelian quantum doubles as an example, we show how the information convex reveals and characterizes deconfined bulk and boundary topological excitations, and the condensation rule relating them. Interference experiments in cold atoms provide potential measurements for the invariant structure of information convex.

cond-mat.str-el↗

arXiv2018

Deciphering the nonlocal entanglement entropy of fracton topological orders

The ground states of topological orders condense extended objects and support topological excitations. This nontrivial property leads to nonzero topological entanglement entropy $S_{topo}$ for conventional topological orders. Fracton topological order is an exotic class of models which is beyond the description of TQFT. With some assumptions about the condensates and the topological excitations, we derive a lower bound of the nonlocal entanglement entropy $S_{nonlocal}$ (a generalization of $S_{topo}$). The lower bound applies to Abelian stabilizer models including conventional topological orders as well as type \Rom{1} and type \Rom{2} fracton models, and it could be used to distinguish them. For fracton models, the lower bound shows that $S_{nonlocal}$ could obtain geometry-dependent values, and $S_{nonlocal}$ is extensive for certain choices of subsystems, including some choices which always give zero for TQFT. The stability of the lower bound under local perturbations is discussed.

cond-mat.str-el↗

arXiv2017

Classification and surface anomaly of glide symmetry protected topological phases in three dimensions

We study glide protected topological (GSPT) phases of interacting bosons and fermions in three spatial dimensions certain on-site symmetries. They are crystalline SPT phases, which are distinguished from a trivial product state only in the presence of non-symmorphic glide symmetry. We classify these GSPT phases with various on-site symmetries such as $U(1)$ and time reversal, and show that they can all be understood by stacking and coupling two-dimensional short-range-entangled phases in a glide-invariant way. Using such a coupled layer construction we study the anomalous surface topological orders of these GSPT phases, which gap out the two-dimensional surface states without breaking any symmetries. This framework is demonstrated in many examples, including the non-symmorphic topological insulator with "hourglass fermion" surface states.

cond-mat.str-el↗

arXiv2015

Basis invariant description of chemical equilibrium with implications for a recent axionic leptogenesis model

We provide a systematic treatment of chemical equilibrium in the presence of a specific type of time dependent background. The type of time dependent background we consider appears, for example, in recently proposed axion/Majoron leptogenesis models [1,2]. In describing the chemical equilibrium we use quantities which are invariant under redefinition of fermion phases (we refer to this redefinition as a change of basis for short), and therefore it is a basis invariant treatment. The change of the anomaly terms due to the change of the path integral measure [3,4] under a basis change is taken into account. We find it is useful to go back and forth between different bases, and there are insights which can be more easily obtained in one basis rather than another. A toy model is provided to illustrate the ideas. For the axion leptogenesis model [1], our result suggests that at $T > 10^{13}$ GeV , when sphaleron processes decouple, and $Γ_{B+L} << H < Γ_L$ (where $H$ is the Hubble parameter at temperature $T$ and $Γ_L$ is the $ΔL = 2$ lepton number violating interaction rate), the amount of $B-L$ created is controlled by the smallness of the sphaleron interaction rate, $Γ_{B+L}$. Therefore it is not as efficient as described. In addition, we notice an interesting modification of gauge boson dispersion relations at subleading order.

hep-ph↗